The skill that knows
how X's algorithm actually works

No generic social media advice. Every recommendation traces to real algorithm behavior from X's open-sourced codebase. Know exactly which signals drive reach.

Get Started View the Algorithm

Installation

Install the X Algorithm Skill for Claude Code with a single command.

npx x-algorithm-skill init

Requires Claude Code. The skill will be added to your Claude Code skills directory.

Usage

Use the skill in Claude Code to get algorithm-backed guidance for your posts.

Get Writing Guidance

View the full algorithm overview and writing guidance.

/x-algorithm

Review a Draft

Get algorithm-backed feedback on your draft with specific improvements.

/x-algorithm review

Here's my draft: [your draft]

Rewrite for Maximum Virality

Get an algorithm-optimized rewrite of your post with maximum engagement potential.

/x-algorithm rewrite

Here's my post: [your post]

Scoring Signals

The algorithm predicts 19 engagement probabilities and combines them into a weighted score. Understanding what drives each signal is key to virality.

final_score = Σ(probability_i × weight_i) + offset

↑ High-Value Signals

Maximize these for maximum reach.

Signal	What It Measures	How to Trigger
Reply	Genuine conversation	Ask questions, invite perspectives
DM Share	Private sharing	Insider info, niche value
Copy Link Share	External sharing	Reference-worthy content
Dwell Time	Time spent reading	Depth, formatting, intrigue
Follow	New follower conversion	Demonstrate ongoing value
Quote	Amplification with commentary	Hot takes worth responding to

→ Medium Signals

Good to have, but lower weight.

Signal	What It Measures
Like	Basic approval
Repost	Amplification
Click	Link/media engagement
Profile Click	Author interest
Photo Expand	Image engagement
Video Quality View	Video completion

↓ Negative Signals

Avoid at all costs.

Signal	What It Causes
Not Interested	Score penalty
Mute Author	Score penalty
Block Author	Severe penalty
Report	Severe penalty

Key Insight

The algorithm doesn't care about "engagement" generically. It cares about specific actions: replies, shares, dwell time. A post that gets 100 likes but no replies scores worse than a post with 50 likes and 20 replies.

Discovery & Retrieval

Before posts can be ranked, they must be retrieved. The algorithm uses a two-tower neural network to efficiently find relevant posts from millions of candidates.

┌─────────────────┐          ┌─────────────────┐
│   USER TOWER    │          │ CANDIDATE TOWER │
│                 │          │                 │
│  User Features  │          │  Post Content   │
│       +         │          │       +         │
│  Engagement     │          │  Author Info    │
│    History      │          │                 │
│       ↓         │          │       ↓         │
│  Transformer    │          │     MLP         │
│       ↓         │          │       ↓         │
│  L2 Normalize   │          │  L2 Normalize   │
└────────┬────────┘          └────────┬────────┘
         │                            │
         └──────────┬─────────────────┘
                    │
            Dot Product Similarity
                    ↓
              Top-K Retrieval

User Tower (Your Audience)

Encodes who a user is based on their features and engagement history. The algorithm looks at up to 128 recent interactions.

Which posts they engaged with
Which authors they engaged with
What actions they took (like, reply, share)
Where they interacted (feed, notifications, search)

Candidate Tower (Your Post)

Encodes each post as a point in high-dimensional space. Posts cluster by topic, style, and author.

Post content embeddings
Author embeddings and characteristics
Processed through 2-layer MLP
L2-normalized to unit sphere

In-Network vs Out-of-Network

In-Network (Thunder)

Posts from accounts the user follows. Sub-millisecond lookup from in-memory store. No retrieval penalty.

Out-of-Network (Phoenix)

Posts from accounts the user doesn't follow. Uses two-tower similarity search. Gets a penalty factor applied.

Practical Takeaways

1. Build consistent engagement patterns. Post consistently in your niche to build a strong author embedding.
2. Earn in-network status. Converting followers is crucial—they see you without the OON penalty.
3. Respect diversity. Don't post 20 times a day. Multiple posts get exponentially discounted.
4. Stay fresh. Posts older than the threshold are filtered out entirely. Timing matters.

Viral Patterns

Proven templates that trigger high-value engagement signals. Each pattern is optimized for specific algorithm behaviors.

The Insight Thread

Dwell time, Follows, Quote tweets

Non-obvious observation as hook, followed by numbered insights and a closing CTA for follow.

View example

I reviewed 200 failed startups for Y Combinator.

Here's what most people get wrong:

1. They pivoted too late, not too early
2. The cofounder conflicts started before the idea
3. "Running out of money" was the symptom, not the cause
4. The winners had customers before they had product

The real lesson: Speed of learning beats everything else.

The Contrarian Take

Replies, Quote tweets, Profile clicks

Bold claim that challenges consensus. Brief explanation without over-explaining.

View example

"Work-life balance" is a trap.

The best founders I know don't balance anything. They integrate.

Work is life. Life is work. The separation is the problem.

The Resource Post

Copy Link Share, DM Share, Saves

Clear promise of value with comprehensive but scannable content.

View example

Complete cold email checklist:

□ Subject line < 6 words
□ First line personalized to them
□ One clear ask
□ Under 100 words total
□ No attachments
□ Send Tuesday-Thursday 9-11am
□ Follow up exactly once at day 3

Save this. Send it to your sales team.

The Curiosity Gap

Dwell time, Click-through, Replies

Create specific curiosity, delay payoff just enough, deliver satisfying answer.

View example

The #1 product at Amazon isn't what you think.

Not AWS. Not Prime. Not Alexa.

It's the internal document culture.

Every major decision starts with a 6-page narrative memo.

That's why they move fast—everyone actually understands the decision.

The Personal Story

Follows, Dwell time, Emotional resonance

Start mid-action (not 'So I was...'), quick context, turning point, and lesson.

View example

"We're shutting down the company."

Hardest email I ever sent. 47 employees. Families. Dreams.

But here's what I didn't expect:

Every single person replied thanking me. Not for the job. For being honest.

Transparency isn't just ethical. It's the only way to leave with your reputation intact.

The Direct Value

Shares, Replies, Practical engagement

Immediate actionable content with no preamble.

View example

Write better emails:

1. Start with the ask
2. Max 5 sentences
3. One action per email

That's it. Stop overthinking.

So You Don't Have To

Share, Follow, Dwell

Establish effort, distill key findings, save them time.

View example

I read 50 books on negotiation so you don't have to.

Here are the 3 things that actually matter:

1. The person who cares less has more power
2. "No" is the start, not the end
3. Silence is your best friend

Everything else is noise.

The Prediction

Quote tweets, Replies, Expert engagement

Specific prediction with timeframe, brief reasoning, optional stake.

View example

Prediction: 50% of SaaS companies will be AI-wrapper businesses by 2026.

Here's why:

- Foundation models are commoditizing
- UX is the only moat
- Infrastructure costs favor thin layers

Bookmark this. Check back in 2 years.

Signal Targeting Summary

Pattern	Primary Signals
Insight Thread	Dwell, Follow, Quote
Contrarian Take	Reply, Quote, Profile Click
Resource Post	Copy Link Share, DM Share
Curiosity Gap	Dwell, Click
Personal Story	Follow, Dwell
Direct Value	Share, Reply
"So You Don't Have To"	Share, Follow, Dwell
Prediction	Quote, Reply

Anti-Patterns: What Gets Penalized

The algorithm has explicit negative signals that reduce your score. Understanding what triggers them is as important as understanding what drives positive engagement.

The Engagement Farmer

Triggers: Not Interested

Explicit calls to like/RT/follow
"Engagement pods" or coordinated liking
Growth hacking tactics

Why suppressed: These signals are learned as inauthentic. The model predicts "not interested" responses.

The Reply Guy

Triggers: Mute

Commenting on every viral post
Self-promotional replies
First-to-respond gaming

Why suppressed: Low-quality engagement patterns lead to mutes.

The Thread Maximalist

Triggers: Algorithm Penalty

Every thought becomes a 20-tweet thread
Padding threads with unnecessary posts
"1/" on content that doesn't need threading

Why suppressed: Author diversity penalty kicks in. Later posts in thread get discounted.

The Controversy Farmer

Triggers: Block, Mute

Intentionally inflammatory takes
Punching down for engagement
Rage bait framing

Why suppressed: Short-term replies, long-term blocks and mutes.

The Copy-Paste Recycler

Triggers: Filtered Out

Same content reposted weekly
Slight variations of the same post
Cross-posting without adaptation

Why suppressed: Deduplication and previously-seen filters catch this.

Red Flags Checklist

Before posting, check:

Am I explicitly asking for engagement? (remove it)
Would I be annoyed seeing this from someone else? (reconsider)
Is this the 5th+ post today? (wait)
Am I being hostile to anyone? (reframe)
Have I posted something similar recently? (don't)
Does this deliver on its hook? (must)
Would someone unfollow over this? (pause)

Writing Like a Human, Not a Bot

New in v1.2

Algorithm optimization means nothing if your posts scream "AI wrote this." Users can smell inauthenticity. The skill now includes comprehensive humanization rules to help you write posts that are both algorithm-optimized AND authentically human.

AI Vocabulary Blacklist

Words that instantly signal AI writing:

leverageutilizecrucialdelvetransformativegroundbreakingunlockpivotalrobustcomprehensive

Structural Tells

Always listing exactly 3 things
"Not just X, but Y" constructions
Em dash overuse for dramatic pauses
"Here's the thing:" formulaic transitions
Perfectly clean structure with no imperfections

Adding Soul

Use contractions naturally (don't, it's, can't)
Vary list lengths (2, 4, or 7 items—not always 3)
Include specific details over vague claims
Allow imperfection—tangents, casual asides
Write like you'd talk to a friend

The full skill includes before/after examples, a complete AI vocabulary blacklist, and a humanization test added to the 6-point audit framework.

Iteration Framework

A systematic approach to improving drafts before posting. Run every draft through this checklist to optimize for the algorithm.

Hook Types Ranked by Signal Strength

Hook Type	Best For	Example Start
Question	Replies	"Why do..."
Contrarian	Quotes, Replies	"[Common belief] is wrong."
Insight	Dwell, Follow	"I discovered..."
Story	Dwell, Follow	"[Dialogue]" or "[Action]"
Resource	Share	"Complete guide to..."
Prediction	Quote, Reply	"Prediction:"

Length Guidelines

Goal	Optimal Length
Maximum reach	71-100 characters
Engagement depth	200-280 characters
Authority building	500-1000 characters
Threads	5-15 posts

The skill that knows how X's algorithm actually works

Installation

Usage

Get Writing Guidance

Review a Draft

Rewrite for Maximum Virality

Scoring Signals

↑ High-Value Signals

→ Medium Signals

↓ Negative Signals

Key Insight

Discovery & Retrieval

User Tower (Your Audience)

Candidate Tower (Your Post)

In-Network vs Out-of-Network

In-Network (Thunder)

Out-of-Network (Phoenix)

Practical Takeaways

Viral Patterns

The Insight Thread

The Contrarian Take

The Resource Post

The Curiosity Gap

The Personal Story

The Direct Value

So You Don't Have To

The Prediction

Signal Targeting Summary

Anti-Patterns: What Gets Penalized

The Engagement Farmer

The Reply Guy

The Thread Maximalist

The Controversy Farmer

The Copy-Paste Recycler

Red Flags Checklist

Writing Like a Human, Not a Bot

AI Vocabulary Blacklist

Structural Tells

Adding Soul

Iteration Framework

Signals

Hook

Structure

Final Check

Hook Types Ranked by Signal Strength

Length Guidelines

The skill that knows
how X's algorithm actually works